Source file src/runtime/stack.go

     1  // Copyright 2013 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  package runtime
     6  
     7  import (
     8  	"internal/abi"
     9  	"internal/cpu"
    10  	"runtime/internal/atomic"
    11  	"runtime/internal/sys"
    12  	"unsafe"
    13  )
    14  
    15  /*
    16  Stack layout parameters.
    17  Included both by runtime (compiled via 6c) and linkers (compiled via gcc).
    18  
    19  The per-goroutine g->stackguard is set to point StackGuard bytes
    20  above the bottom of the stack.  Each function compares its stack
    21  pointer against g->stackguard to check for overflow.  To cut one
    22  instruction from the check sequence for functions with tiny frames,
    23  the stack is allowed to protrude StackSmall bytes below the stack
    24  guard.  Functions with large frames don't bother with the check and
    25  always call morestack.  The sequences are (for amd64, others are
    26  similar):
    27  
    28  	guard = g->stackguard
    29  	frame = function's stack frame size
    30  	argsize = size of function arguments (call + return)
    31  
    32  	stack frame size <= StackSmall:
    33  		CMPQ guard, SP
    34  		JHI 3(PC)
    35  		MOVQ m->morearg, $(argsize << 32)
    36  		CALL morestack(SB)
    37  
    38  	stack frame size > StackSmall but < StackBig
    39  		LEAQ (frame-StackSmall)(SP), R0
    40  		CMPQ guard, R0
    41  		JHI 3(PC)
    42  		MOVQ m->morearg, $(argsize << 32)
    43  		CALL morestack(SB)
    44  
    45  	stack frame size >= StackBig:
    46  		MOVQ m->morearg, $((argsize << 32) | frame)
    47  		CALL morestack(SB)
    48  
    49  The bottom StackGuard - StackSmall bytes are important: there has
    50  to be enough room to execute functions that refuse to check for
    51  stack overflow, either because they need to be adjacent to the
    52  actual caller's frame (deferproc) or because they handle the imminent
    53  stack overflow (morestack).
    54  
    55  For example, deferproc might call malloc, which does one of the
    56  above checks (without allocating a full frame), which might trigger
    57  a call to morestack.  This sequence needs to fit in the bottom
    58  section of the stack.  On amd64, morestack's frame is 40 bytes, and
    59  deferproc's frame is 56 bytes.  That fits well within the
    60  StackGuard - StackSmall bytes at the bottom.
    61  The linkers explore all possible call traces involving non-splitting
    62  functions to make sure that this limit cannot be violated.
    63  */
    64  
    65  const (
    66  	// StackSystem is a number of additional bytes to add
    67  	// to each stack below the usual guard area for OS-specific
    68  	// purposes like signal handling. Used on Windows, Plan 9,
    69  	// and iOS because they do not use a separate stack.
    70  	_StackSystem = sys.GoosWindows*512*sys.PtrSize + sys.GoosPlan9*512 + sys.GoosIos*sys.GoarchArm64*1024
    71  
    72  	// The minimum size of stack used by Go code
    73  	_StackMin = 2048
    74  
    75  	// The minimum stack size to allocate.
    76  	// The hackery here rounds FixedStack0 up to a power of 2.
    77  	_FixedStack0 = _StackMin + _StackSystem
    78  	_FixedStack1 = _FixedStack0 - 1
    79  	_FixedStack2 = _FixedStack1 | (_FixedStack1 >> 1)
    80  	_FixedStack3 = _FixedStack2 | (_FixedStack2 >> 2)
    81  	_FixedStack4 = _FixedStack3 | (_FixedStack3 >> 4)
    82  	_FixedStack5 = _FixedStack4 | (_FixedStack4 >> 8)
    83  	_FixedStack6 = _FixedStack5 | (_FixedStack5 >> 16)
    84  	_FixedStack  = _FixedStack6 + 1
    85  
    86  	// Functions that need frames bigger than this use an extra
    87  	// instruction to do the stack split check, to avoid overflow
    88  	// in case SP - framesize wraps below zero.
    89  	// This value can be no bigger than the size of the unmapped
    90  	// space at zero.
    91  	_StackBig = 4096
    92  
    93  	// The stack guard is a pointer this many bytes above the
    94  	// bottom of the stack.
    95  	//
    96  	// The guard leaves enough room for one _StackSmall frame plus
    97  	// a _StackLimit chain of NOSPLIT calls plus _StackSystem
    98  	// bytes for the OS.
    99  	_StackGuard = 928*sys.StackGuardMultiplier + _StackSystem
   100  
   101  	// After a stack split check the SP is allowed to be this
   102  	// many bytes below the stack guard. This saves an instruction
   103  	// in the checking sequence for tiny frames.
   104  	_StackSmall = 128
   105  
   106  	// The maximum number of bytes that a chain of NOSPLIT
   107  	// functions can use.
   108  	_StackLimit = _StackGuard - _StackSystem - _StackSmall
   109  )
   110  
   111  const (
   112  	// stackDebug == 0: no logging
   113  	//            == 1: logging of per-stack operations
   114  	//            == 2: logging of per-frame operations
   115  	//            == 3: logging of per-word updates
   116  	//            == 4: logging of per-word reads
   117  	stackDebug       = 0
   118  	stackFromSystem  = 0 // allocate stacks from system memory instead of the heap
   119  	stackFaultOnFree = 0 // old stacks are mapped noaccess to detect use after free
   120  	stackPoisonCopy  = 0 // fill stack that should not be accessed with garbage, to detect bad dereferences during copy
   121  	stackNoCache     = 0 // disable per-P small stack caches
   122  
   123  	// check the BP links during traceback.
   124  	debugCheckBP = false
   125  )
   126  
   127  const (
   128  	uintptrMask = 1<<(8*sys.PtrSize) - 1
   129  
   130  	// The values below can be stored to g.stackguard0 to force
   131  	// the next stack check to fail.
   132  	// These are all larger than any real SP.
   133  
   134  	// Goroutine preemption request.
   135  	// 0xfffffade in hex.
   136  	stackPreempt = uintptrMask & -1314
   137  
   138  	// Thread is forking. Causes a split stack check failure.
   139  	// 0xfffffb2e in hex.
   140  	stackFork = uintptrMask & -1234
   141  
   142  	// Force a stack movement. Used for debugging.
   143  	// 0xfffffeed in hex.
   144  	stackForceMove = uintptrMask & -275
   145  )
   146  
   147  // Global pool of spans that have free stacks.
   148  // Stacks are assigned an order according to size.
   149  //     order = log_2(size/FixedStack)
   150  // There is a free list for each order.
   151  var stackpool [_NumStackOrders]struct {
   152  	item stackpoolItem
   153  	_    [cpu.CacheLinePadSize - unsafe.Sizeof(stackpoolItem{})%cpu.CacheLinePadSize]byte
   154  }
   155  
   156  //go:notinheap
   157  type stackpoolItem struct {
   158  	mu   mutex
   159  	span mSpanList
   160  }
   161  
   162  // Global pool of large stack spans.
   163  var stackLarge struct {
   164  	lock mutex
   165  	free [heapAddrBits - pageShift]mSpanList // free lists by log_2(s.npages)
   166  }
   167  
   168  func stackinit() {
   169  	if _StackCacheSize&_PageMask != 0 {
   170  		throw("cache size must be a multiple of page size")
   171  	}
   172  	for i := range stackpool {
   173  		stackpool[i].item.span.init()
   174  		lockInit(&stackpool[i].item.mu, lockRankStackpool)
   175  	}
   176  	for i := range stackLarge.free {
   177  		stackLarge.free[i].init()
   178  		lockInit(&stackLarge.lock, lockRankStackLarge)
   179  	}
   180  }
   181  
   182  // stacklog2 returns ⌊log_2(n)⌋.
   183  func stacklog2(n uintptr) int {
   184  	log2 := 0
   185  	for n > 1 {
   186  		n >>= 1
   187  		log2++
   188  	}
   189  	return log2
   190  }
   191  
   192  // Allocates a stack from the free pool. Must be called with
   193  // stackpool[order].item.mu held.
   194  func stackpoolalloc(order uint8) gclinkptr {
   195  	list := &stackpool[order].item.span
   196  	s := list.first
   197  	lockWithRankMayAcquire(&mheap_.lock, lockRankMheap)
   198  	if s == nil {
   199  		// no free stacks. Allocate another span worth.
   200  		s = mheap_.allocManual(_StackCacheSize>>_PageShift, spanAllocStack)
   201  		if s == nil {
   202  			throw("out of memory")
   203  		}
   204  		if s.allocCount != 0 {
   205  			throw("bad allocCount")
   206  		}
   207  		if s.manualFreeList.ptr() != nil {
   208  			throw("bad manualFreeList")
   209  		}
   210  		osStackAlloc(s)
   211  		s.elemsize = _FixedStack << order
   212  		for i := uintptr(0); i < _StackCacheSize; i += s.elemsize {
   213  			x := gclinkptr(s.base() + i)
   214  			x.ptr().next = s.manualFreeList
   215  			s.manualFreeList = x
   216  		}
   217  		list.insert(s)
   218  	}
   219  	x := s.manualFreeList
   220  	if x.ptr() == nil {
   221  		throw("span has no free stacks")
   222  	}
   223  	s.manualFreeList = x.ptr().next
   224  	s.allocCount++
   225  	if s.manualFreeList.ptr() == nil {
   226  		// all stacks in s are allocated.
   227  		list.remove(s)
   228  	}
   229  	return x
   230  }
   231  
   232  // Adds stack x to the free pool. Must be called with stackpool[order].item.mu held.
   233  func stackpoolfree(x gclinkptr, order uint8) {
   234  	s := spanOfUnchecked(uintptr(x))
   235  	if s.state.get() != mSpanManual {
   236  		throw("freeing stack not in a stack span")
   237  	}
   238  	if s.manualFreeList.ptr() == nil {
   239  		// s will now have a free stack
   240  		stackpool[order].item.span.insert(s)
   241  	}
   242  	x.ptr().next = s.manualFreeList
   243  	s.manualFreeList = x
   244  	s.allocCount--
   245  	if gcphase == _GCoff && s.allocCount == 0 {
   246  		// Span is completely free. Return it to the heap
   247  		// immediately if we're sweeping.
   248  		//
   249  		// If GC is active, we delay the free until the end of
   250  		// GC to avoid the following type of situation:
   251  		//
   252  		// 1) GC starts, scans a SudoG but does not yet mark the SudoG.elem pointer
   253  		// 2) The stack that pointer points to is copied
   254  		// 3) The old stack is freed
   255  		// 4) The containing span is marked free
   256  		// 5) GC attempts to mark the SudoG.elem pointer. The
   257  		//    marking fails because the pointer looks like a
   258  		//    pointer into a free span.
   259  		//
   260  		// By not freeing, we prevent step #4 until GC is done.
   261  		stackpool[order].item.span.remove(s)
   262  		s.manualFreeList = 0
   263  		osStackFree(s)
   264  		mheap_.freeManual(s, spanAllocStack)
   265  	}
   266  }
   267  
   268  // stackcacherefill/stackcacherelease implement a global pool of stack segments.
   269  // The pool is required to prevent unlimited growth of per-thread caches.
   270  //
   271  //go:systemstack
   272  func stackcacherefill(c *mcache, order uint8) {
   273  	if stackDebug >= 1 {
   274  		print("stackcacherefill order=", order, "\n")
   275  	}
   276  
   277  	// Grab some stacks from the global cache.
   278  	// Grab half of the allowed capacity (to prevent thrashing).
   279  	var list gclinkptr
   280  	var size uintptr
   281  	lock(&stackpool[order].item.mu)
   282  	for size < _StackCacheSize/2 {
   283  		x := stackpoolalloc(order)
   284  		x.ptr().next = list
   285  		list = x
   286  		size += _FixedStack << order
   287  	}
   288  	unlock(&stackpool[order].item.mu)
   289  	c.stackcache[order].list = list
   290  	c.stackcache[order].size = size
   291  }
   292  
   293  //go:systemstack
   294  func stackcacherelease(c *mcache, order uint8) {
   295  	if stackDebug >= 1 {
   296  		print("stackcacherelease order=", order, "\n")
   297  	}
   298  	x := c.stackcache[order].list
   299  	size := c.stackcache[order].size
   300  	lock(&stackpool[order].item.mu)
   301  	for size > _StackCacheSize/2 {
   302  		y := x.ptr().next
   303  		stackpoolfree(x, order)
   304  		x = y
   305  		size -= _FixedStack << order
   306  	}
   307  	unlock(&stackpool[order].item.mu)
   308  	c.stackcache[order].list = x
   309  	c.stackcache[order].size = size
   310  }
   311  
   312  //go:systemstack
   313  func stackcache_clear(c *mcache) {
   314  	if stackDebug >= 1 {
   315  		print("stackcache clear\n")
   316  	}
   317  	for order := uint8(0); order < _NumStackOrders; order++ {
   318  		lock(&stackpool[order].item.mu)
   319  		x := c.stackcache[order].list
   320  		for x.ptr() != nil {
   321  			y := x.ptr().next
   322  			stackpoolfree(x, order)
   323  			x = y
   324  		}
   325  		c.stackcache[order].list = 0
   326  		c.stackcache[order].size = 0
   327  		unlock(&stackpool[order].item.mu)
   328  	}
   329  }
   330  
   331  // stackalloc allocates an n byte stack.
   332  //
   333  // stackalloc must run on the system stack because it uses per-P
   334  // resources and must not split the stack.
   335  //
   336  //go:systemstack
   337  func stackalloc(n uint32) stack {
   338  	// Stackalloc must be called on scheduler stack, so that we
   339  	// never try to grow the stack during the code that stackalloc runs.
   340  	// Doing so would cause a deadlock (issue 1547).
   341  	thisg := getg()
   342  	if thisg != thisg.m.g0 {
   343  		throw("stackalloc not on scheduler stack")
   344  	}
   345  	if n&(n-1) != 0 {
   346  		throw("stack size not a power of 2")
   347  	}
   348  	if stackDebug >= 1 {
   349  		print("stackalloc ", n, "\n")
   350  	}
   351  
   352  	if debug.efence != 0 || stackFromSystem != 0 {
   353  		n = uint32(alignUp(uintptr(n), physPageSize))
   354  		v := sysAlloc(uintptr(n), &memstats.stacks_sys)
   355  		if v == nil {
   356  			throw("out of memory (stackalloc)")
   357  		}
   358  		return stack{uintptr(v), uintptr(v) + uintptr(n)}
   359  	}
   360  
   361  	// Small stacks are allocated with a fixed-size free-list allocator.
   362  	// If we need a stack of a bigger size, we fall back on allocating
   363  	// a dedicated span.
   364  	var v unsafe.Pointer
   365  	if n < _FixedStack<<_NumStackOrders && n < _StackCacheSize {
   366  		order := uint8(0)
   367  		n2 := n
   368  		for n2 > _FixedStack {
   369  			order++
   370  			n2 >>= 1
   371  		}
   372  		var x gclinkptr
   373  		if stackNoCache != 0 || thisg.m.p == 0 || thisg.m.preemptoff != "" {
   374  			// thisg.m.p == 0 can happen in the guts of exitsyscall
   375  			// or procresize. Just get a stack from the global pool.
   376  			// Also don't touch stackcache during gc
   377  			// as it's flushed concurrently.
   378  			lock(&stackpool[order].item.mu)
   379  			x = stackpoolalloc(order)
   380  			unlock(&stackpool[order].item.mu)
   381  		} else {
   382  			c := thisg.m.p.ptr().mcache
   383  			x = c.stackcache[order].list
   384  			if x.ptr() == nil {
   385  				stackcacherefill(c, order)
   386  				x = c.stackcache[order].list
   387  			}
   388  			c.stackcache[order].list = x.ptr().next
   389  			c.stackcache[order].size -= uintptr(n)
   390  		}
   391  		v = unsafe.Pointer(x)
   392  	} else {
   393  		var s *mspan
   394  		npage := uintptr(n) >> _PageShift
   395  		log2npage := stacklog2(npage)
   396  
   397  		// Try to get a stack from the large stack cache.
   398  		lock(&stackLarge.lock)
   399  		if !stackLarge.free[log2npage].isEmpty() {
   400  			s = stackLarge.free[log2npage].first
   401  			stackLarge.free[log2npage].remove(s)
   402  		}
   403  		unlock(&stackLarge.lock)
   404  
   405  		lockWithRankMayAcquire(&mheap_.lock, lockRankMheap)
   406  
   407  		if s == nil {
   408  			// Allocate a new stack from the heap.
   409  			s = mheap_.allocManual(npage, spanAllocStack)
   410  			if s == nil {
   411  				throw("out of memory")
   412  			}
   413  			osStackAlloc(s)
   414  			s.elemsize = uintptr(n)
   415  		}
   416  		v = unsafe.Pointer(s.base())
   417  	}
   418  
   419  	if raceenabled {
   420  		racemalloc(v, uintptr(n))
   421  	}
   422  	if msanenabled {
   423  		msanmalloc(v, uintptr(n))
   424  	}
   425  	if stackDebug >= 1 {
   426  		print("  allocated ", v, "\n")
   427  	}
   428  	return stack{uintptr(v), uintptr(v) + uintptr(n)}
   429  }
   430  
   431  // stackfree frees an n byte stack allocation at stk.
   432  //
   433  // stackfree must run on the system stack because it uses per-P
   434  // resources and must not split the stack.
   435  //
   436  //go:systemstack
   437  func stackfree(stk stack) {
   438  	gp := getg()
   439  	v := unsafe.Pointer(stk.lo)
   440  	n := stk.hi - stk.lo
   441  	if n&(n-1) != 0 {
   442  		throw("stack not a power of 2")
   443  	}
   444  	if stk.lo+n < stk.hi {
   445  		throw("bad stack size")
   446  	}
   447  	if stackDebug >= 1 {
   448  		println("stackfree", v, n)
   449  		memclrNoHeapPointers(v, n) // for testing, clobber stack data
   450  	}
   451  	if debug.efence != 0 || stackFromSystem != 0 {
   452  		if debug.efence != 0 || stackFaultOnFree != 0 {
   453  			sysFault(v, n)
   454  		} else {
   455  			sysFree(v, n, &memstats.stacks_sys)
   456  		}
   457  		return
   458  	}
   459  	if msanenabled {
   460  		msanfree(v, n)
   461  	}
   462  	if n < _FixedStack<<_NumStackOrders && n < _StackCacheSize {
   463  		order := uint8(0)
   464  		n2 := n
   465  		for n2 > _FixedStack {
   466  			order++
   467  			n2 >>= 1
   468  		}
   469  		x := gclinkptr(v)
   470  		if stackNoCache != 0 || gp.m.p == 0 || gp.m.preemptoff != "" {
   471  			lock(&stackpool[order].item.mu)
   472  			stackpoolfree(x, order)
   473  			unlock(&stackpool[order].item.mu)
   474  		} else {
   475  			c := gp.m.p.ptr().mcache
   476  			if c.stackcache[order].size >= _StackCacheSize {
   477  				stackcacherelease(c, order)
   478  			}
   479  			x.ptr().next = c.stackcache[order].list
   480  			c.stackcache[order].list = x
   481  			c.stackcache[order].size += n
   482  		}
   483  	} else {
   484  		s := spanOfUnchecked(uintptr(v))
   485  		if s.state.get() != mSpanManual {
   486  			println(hex(s.base()), v)
   487  			throw("bad span state")
   488  		}
   489  		if gcphase == _GCoff {
   490  			// Free the stack immediately if we're
   491  			// sweeping.
   492  			osStackFree(s)
   493  			mheap_.freeManual(s, spanAllocStack)
   494  		} else {
   495  			// If the GC is running, we can't return a
   496  			// stack span to the heap because it could be
   497  			// reused as a heap span, and this state
   498  			// change would race with GC. Add it to the
   499  			// large stack cache instead.
   500  			log2npage := stacklog2(s.npages)
   501  			lock(&stackLarge.lock)
   502  			stackLarge.free[log2npage].insert(s)
   503  			unlock(&stackLarge.lock)
   504  		}
   505  	}
   506  }
   507  
   508  var maxstacksize uintptr = 1 << 20 // enough until runtime.main sets it for real
   509  
   510  var maxstackceiling = maxstacksize
   511  
   512  var ptrnames = []string{
   513  	0: "scalar",
   514  	1: "ptr",
   515  }
   516  
   517  // Stack frame layout
   518  //
   519  // (x86)
   520  // +------------------+
   521  // | args from caller |
   522  // +------------------+ <- frame->argp
   523  // |  return address  |
   524  // +------------------+
   525  // |  caller's BP (*) | (*) if framepointer_enabled && varp < sp
   526  // +------------------+ <- frame->varp
   527  // |     locals       |
   528  // +------------------+
   529  // |  args to callee  |
   530  // +------------------+ <- frame->sp
   531  //
   532  // (arm)
   533  // +------------------+
   534  // | args from caller |
   535  // +------------------+ <- frame->argp
   536  // | caller's retaddr |
   537  // +------------------+ <- frame->varp
   538  // |     locals       |
   539  // +------------------+
   540  // |  args to callee  |
   541  // +------------------+
   542  // |  return address  |
   543  // +------------------+ <- frame->sp
   544  
   545  type adjustinfo struct {
   546  	old   stack
   547  	delta uintptr // ptr distance from old to new stack (newbase - oldbase)
   548  	cache pcvalueCache
   549  
   550  	// sghi is the highest sudog.elem on the stack.
   551  	sghi uintptr
   552  }
   553  
   554  // Adjustpointer checks whether *vpp is in the old stack described by adjinfo.
   555  // If so, it rewrites *vpp to point into the new stack.
   556  func adjustpointer(adjinfo *adjustinfo, vpp unsafe.Pointer) {
   557  	pp := (*uintptr)(vpp)
   558  	p := *pp
   559  	if stackDebug >= 4 {
   560  		print("        ", pp, ":", hex(p), "\n")
   561  	}
   562  	if adjinfo.old.lo <= p && p < adjinfo.old.hi {
   563  		*pp = p + adjinfo.delta
   564  		if stackDebug >= 3 {
   565  			print("        adjust ptr ", pp, ":", hex(p), " -> ", hex(*pp), "\n")
   566  		}
   567  	}
   568  }
   569  
   570  // Information from the compiler about the layout of stack frames.
   571  // Note: this type must agree with reflect.bitVector.
   572  type bitvector struct {
   573  	n        int32 // # of bits
   574  	bytedata *uint8
   575  }
   576  
   577  // ptrbit returns the i'th bit in bv.
   578  // ptrbit is less efficient than iterating directly over bitvector bits,
   579  // and should only be used in non-performance-critical code.
   580  // See adjustpointers for an example of a high-efficiency walk of a bitvector.
   581  func (bv *bitvector) ptrbit(i uintptr) uint8 {
   582  	b := *(addb(bv.bytedata, i/8))
   583  	return (b >> (i % 8)) & 1
   584  }
   585  
   586  // bv describes the memory starting at address scanp.
   587  // Adjust any pointers contained therein.
   588  func adjustpointers(scanp unsafe.Pointer, bv *bitvector, adjinfo *adjustinfo, f funcInfo) {
   589  	minp := adjinfo.old.lo
   590  	maxp := adjinfo.old.hi
   591  	delta := adjinfo.delta
   592  	num := uintptr(bv.n)
   593  	// If this frame might contain channel receive slots, use CAS
   594  	// to adjust pointers. If the slot hasn't been received into
   595  	// yet, it may contain stack pointers and a concurrent send
   596  	// could race with adjusting those pointers. (The sent value
   597  	// itself can never contain stack pointers.)
   598  	useCAS := uintptr(scanp) < adjinfo.sghi
   599  	for i := uintptr(0); i < num; i += 8 {
   600  		if stackDebug >= 4 {
   601  			for j := uintptr(0); j < 8; j++ {
   602  				print("        ", add(scanp, (i+j)*sys.PtrSize), ":", ptrnames[bv.ptrbit(i+j)], ":", hex(*(*uintptr)(add(scanp, (i+j)*sys.PtrSize))), " # ", i, " ", *addb(bv.bytedata, i/8), "\n")
   603  			}
   604  		}
   605  		b := *(addb(bv.bytedata, i/8))
   606  		for b != 0 {
   607  			j := uintptr(sys.Ctz8(b))
   608  			b &= b - 1
   609  			pp := (*uintptr)(add(scanp, (i+j)*sys.PtrSize))
   610  		retry:
   611  			p := *pp
   612  			if f.valid() && 0 < p && p < minLegalPointer && debug.invalidptr != 0 {
   613  				// Looks like a junk value in a pointer slot.
   614  				// Live analysis wrong?
   615  				getg().m.traceback = 2
   616  				print("runtime: bad pointer in frame ", funcname(f), " at ", pp, ": ", hex(p), "\n")
   617  				throw("invalid pointer found on stack")
   618  			}
   619  			if minp <= p && p < maxp {
   620  				if stackDebug >= 3 {
   621  					print("adjust ptr ", hex(p), " ", funcname(f), "\n")
   622  				}
   623  				if useCAS {
   624  					ppu := (*unsafe.Pointer)(unsafe.Pointer(pp))
   625  					if !atomic.Casp1(ppu, unsafe.Pointer(p), unsafe.Pointer(p+delta)) {
   626  						goto retry
   627  					}
   628  				} else {
   629  					*pp = p + delta
   630  				}
   631  			}
   632  		}
   633  	}
   634  }
   635  
   636  // Note: the argument/return area is adjusted by the callee.
   637  func adjustframe(frame *stkframe, arg unsafe.Pointer) bool {
   638  	adjinfo := (*adjustinfo)(arg)
   639  	if frame.continpc == 0 {
   640  		// Frame is dead.
   641  		return true
   642  	}
   643  	f := frame.fn
   644  	if stackDebug >= 2 {
   645  		print("    adjusting ", funcname(f), " frame=[", hex(frame.sp), ",", hex(frame.fp), "] pc=", hex(frame.pc), " continpc=", hex(frame.continpc), "\n")
   646  	}
   647  	if f.funcID == funcID_systemstack_switch {
   648  		// A special routine at the bottom of stack of a goroutine that does a systemstack call.
   649  		// We will allow it to be copied even though we don't
   650  		// have full GC info for it (because it is written in asm).
   651  		return true
   652  	}
   653  
   654  	locals, args, objs := getStackMap(frame, &adjinfo.cache, true)
   655  
   656  	// Adjust local variables if stack frame has been allocated.
   657  	if locals.n > 0 {
   658  		size := uintptr(locals.n) * sys.PtrSize
   659  		adjustpointers(unsafe.Pointer(frame.varp-size), &locals, adjinfo, f)
   660  	}
   661  
   662  	// Adjust saved base pointer if there is one.
   663  	// TODO what about arm64 frame pointer adjustment?
   664  	if sys.ArchFamily == sys.AMD64 && frame.argp-frame.varp == 2*sys.PtrSize {
   665  		if stackDebug >= 3 {
   666  			print("      saved bp\n")
   667  		}
   668  		if debugCheckBP {
   669  			// Frame pointers should always point to the next higher frame on
   670  			// the Go stack (or be nil, for the top frame on the stack).
   671  			bp := *(*uintptr)(unsafe.Pointer(frame.varp))
   672  			if bp != 0 && (bp < adjinfo.old.lo || bp >= adjinfo.old.hi) {
   673  				println("runtime: found invalid frame pointer")
   674  				print("bp=", hex(bp), " min=", hex(adjinfo.old.lo), " max=", hex(adjinfo.old.hi), "\n")
   675  				throw("bad frame pointer")
   676  			}
   677  		}
   678  		adjustpointer(adjinfo, unsafe.Pointer(frame.varp))
   679  	}
   680  
   681  	// Adjust arguments.
   682  	if args.n > 0 {
   683  		if stackDebug >= 3 {
   684  			print("      args\n")
   685  		}
   686  		adjustpointers(unsafe.Pointer(frame.argp), &args, adjinfo, funcInfo{})
   687  	}
   688  
   689  	// Adjust pointers in all stack objects (whether they are live or not).
   690  	// See comments in mgcmark.go:scanframeworker.
   691  	if frame.varp != 0 {
   692  		for _, obj := range objs {
   693  			off := obj.off
   694  			base := frame.varp // locals base pointer
   695  			if off >= 0 {
   696  				base = frame.argp // arguments and return values base pointer
   697  			}
   698  			p := base + uintptr(off)
   699  			if p < frame.sp {
   700  				// Object hasn't been allocated in the frame yet.
   701  				// (Happens when the stack bounds check fails and
   702  				// we call into morestack.)
   703  				continue
   704  			}
   705  			ptrdata := obj.ptrdata()
   706  			gcdata := obj.gcdata
   707  			var s *mspan
   708  			if obj.useGCProg() {
   709  				// See comments in mgcmark.go:scanstack
   710  				s = materializeGCProg(ptrdata, gcdata)
   711  				gcdata = (*byte)(unsafe.Pointer(s.startAddr))
   712  			}
   713  			for i := uintptr(0); i < ptrdata; i += sys.PtrSize {
   714  				if *addb(gcdata, i/(8*sys.PtrSize))>>(i/sys.PtrSize&7)&1 != 0 {
   715  					adjustpointer(adjinfo, unsafe.Pointer(p+i))
   716  				}
   717  			}
   718  			if s != nil {
   719  				dematerializeGCProg(s)
   720  			}
   721  		}
   722  	}
   723  
   724  	return true
   725  }
   726  
   727  func adjustctxt(gp *g, adjinfo *adjustinfo) {
   728  	adjustpointer(adjinfo, unsafe.Pointer(&gp.sched.ctxt))
   729  	if !framepointer_enabled {
   730  		return
   731  	}
   732  	if debugCheckBP {
   733  		bp := gp.sched.bp
   734  		if bp != 0 && (bp < adjinfo.old.lo || bp >= adjinfo.old.hi) {
   735  			println("runtime: found invalid top frame pointer")
   736  			print("bp=", hex(bp), " min=", hex(adjinfo.old.lo), " max=", hex(adjinfo.old.hi), "\n")
   737  			throw("bad top frame pointer")
   738  		}
   739  	}
   740  	adjustpointer(adjinfo, unsafe.Pointer(&gp.sched.bp))
   741  }
   742  
   743  func adjustdefers(gp *g, adjinfo *adjustinfo) {
   744  	// Adjust pointers in the Defer structs.
   745  	// We need to do this first because we need to adjust the
   746  	// defer.link fields so we always work on the new stack.
   747  	adjustpointer(adjinfo, unsafe.Pointer(&gp._defer))
   748  	for d := gp._defer; d != nil; d = d.link {
   749  		adjustpointer(adjinfo, unsafe.Pointer(&d.fn))
   750  		adjustpointer(adjinfo, unsafe.Pointer(&d.sp))
   751  		adjustpointer(adjinfo, unsafe.Pointer(&d._panic))
   752  		adjustpointer(adjinfo, unsafe.Pointer(&d.link))
   753  		adjustpointer(adjinfo, unsafe.Pointer(&d.varp))
   754  		adjustpointer(adjinfo, unsafe.Pointer(&d.fd))
   755  	}
   756  
   757  	// Adjust defer argument blocks the same way we adjust active stack frames.
   758  	// Note: this code is after the loop above, so that if a defer record is
   759  	// stack allocated, we work on the copy in the new stack.
   760  	tracebackdefers(gp, adjustframe, noescape(unsafe.Pointer(adjinfo)))
   761  }
   762  
   763  func adjustpanics(gp *g, adjinfo *adjustinfo) {
   764  	// Panics are on stack and already adjusted.
   765  	// Update pointer to head of list in G.
   766  	adjustpointer(adjinfo, unsafe.Pointer(&gp._panic))
   767  }
   768  
   769  func adjustsudogs(gp *g, adjinfo *adjustinfo) {
   770  	// the data elements pointed to by a SudoG structure
   771  	// might be in the stack.
   772  	for s := gp.waiting; s != nil; s = s.waitlink {
   773  		adjustpointer(adjinfo, unsafe.Pointer(&s.elem))
   774  	}
   775  }
   776  
   777  func fillstack(stk stack, b byte) {
   778  	for p := stk.lo; p < stk.hi; p++ {
   779  		*(*byte)(unsafe.Pointer(p)) = b
   780  	}
   781  }
   782  
   783  func findsghi(gp *g, stk stack) uintptr {
   784  	var sghi uintptr
   785  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
   786  		p := uintptr(sg.elem) + uintptr(sg.c.elemsize)
   787  		if stk.lo <= p && p < stk.hi && p > sghi {
   788  			sghi = p
   789  		}
   790  	}
   791  	return sghi
   792  }
   793  
   794  // syncadjustsudogs adjusts gp's sudogs and copies the part of gp's
   795  // stack they refer to while synchronizing with concurrent channel
   796  // operations. It returns the number of bytes of stack copied.
   797  func syncadjustsudogs(gp *g, used uintptr, adjinfo *adjustinfo) uintptr {
   798  	if gp.waiting == nil {
   799  		return 0
   800  	}
   801  
   802  	// Lock channels to prevent concurrent send/receive.
   803  	var lastc *hchan
   804  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
   805  		if sg.c != lastc {
   806  			// There is a ranking cycle here between gscan bit and
   807  			// hchan locks. Normally, we only allow acquiring hchan
   808  			// locks and then getting a gscan bit. In this case, we
   809  			// already have the gscan bit. We allow acquiring hchan
   810  			// locks here as a special case, since a deadlock can't
   811  			// happen because the G involved must already be
   812  			// suspended. So, we get a special hchan lock rank here
   813  			// that is lower than gscan, but doesn't allow acquiring
   814  			// any other locks other than hchan.
   815  			lockWithRank(&sg.c.lock, lockRankHchanLeaf)
   816  		}
   817  		lastc = sg.c
   818  	}
   819  
   820  	// Adjust sudogs.
   821  	adjustsudogs(gp, adjinfo)
   822  
   823  	// Copy the part of the stack the sudogs point in to
   824  	// while holding the lock to prevent races on
   825  	// send/receive slots.
   826  	var sgsize uintptr
   827  	if adjinfo.sghi != 0 {
   828  		oldBot := adjinfo.old.hi - used
   829  		newBot := oldBot + adjinfo.delta
   830  		sgsize = adjinfo.sghi - oldBot
   831  		memmove(unsafe.Pointer(newBot), unsafe.Pointer(oldBot), sgsize)
   832  	}
   833  
   834  	// Unlock channels.
   835  	lastc = nil
   836  	for sg := gp.waiting; sg != nil; sg = sg.waitlink {
   837  		if sg.c != lastc {
   838  			unlock(&sg.c.lock)
   839  		}
   840  		lastc = sg.c
   841  	}
   842  
   843  	return sgsize
   844  }
   845  
   846  // Copies gp's stack to a new stack of a different size.
   847  // Caller must have changed gp status to Gcopystack.
   848  func copystack(gp *g, newsize uintptr) {
   849  	if gp.syscallsp != 0 {
   850  		throw("stack growth not allowed in system call")
   851  	}
   852  	old := gp.stack
   853  	if old.lo == 0 {
   854  		throw("nil stackbase")
   855  	}
   856  	used := old.hi - gp.sched.sp
   857  
   858  	// allocate new stack
   859  	new := stackalloc(uint32(newsize))
   860  	if stackPoisonCopy != 0 {
   861  		fillstack(new, 0xfd)
   862  	}
   863  	if stackDebug >= 1 {
   864  		print("copystack gp=", gp, " [", hex(old.lo), " ", hex(old.hi-used), " ", hex(old.hi), "]", " -> [", hex(new.lo), " ", hex(new.hi-used), " ", hex(new.hi), "]/", newsize, "\n")
   865  	}
   866  
   867  	// Compute adjustment.
   868  	var adjinfo adjustinfo
   869  	adjinfo.old = old
   870  	adjinfo.delta = new.hi - old.hi
   871  
   872  	// Adjust sudogs, synchronizing with channel ops if necessary.
   873  	ncopy := used
   874  	if !gp.activeStackChans {
   875  		if newsize < old.hi-old.lo && atomic.Load8(&gp.parkingOnChan) != 0 {
   876  			// It's not safe for someone to shrink this stack while we're actively
   877  			// parking on a channel, but it is safe to grow since we do that
   878  			// ourselves and explicitly don't want to synchronize with channels
   879  			// since we could self-deadlock.
   880  			throw("racy sudog adjustment due to parking on channel")
   881  		}
   882  		adjustsudogs(gp, &adjinfo)
   883  	} else {
   884  		// sudogs may be pointing in to the stack and gp has
   885  		// released channel locks, so other goroutines could
   886  		// be writing to gp's stack. Find the highest such
   887  		// pointer so we can handle everything there and below
   888  		// carefully. (This shouldn't be far from the bottom
   889  		// of the stack, so there's little cost in handling
   890  		// everything below it carefully.)
   891  		adjinfo.sghi = findsghi(gp, old)
   892  
   893  		// Synchronize with channel ops and copy the part of
   894  		// the stack they may interact with.
   895  		ncopy -= syncadjustsudogs(gp, used, &adjinfo)
   896  	}
   897  
   898  	// Copy the stack (or the rest of it) to the new location
   899  	memmove(unsafe.Pointer(new.hi-ncopy), unsafe.Pointer(old.hi-ncopy), ncopy)
   900  
   901  	// Adjust remaining structures that have pointers into stacks.
   902  	// We have to do most of these before we traceback the new
   903  	// stack because gentraceback uses them.
   904  	adjustctxt(gp, &adjinfo)
   905  	adjustdefers(gp, &adjinfo)
   906  	adjustpanics(gp, &adjinfo)
   907  	if adjinfo.sghi != 0 {
   908  		adjinfo.sghi += adjinfo.delta
   909  	}
   910  
   911  	// Swap out old stack for new one
   912  	gp.stack = new
   913  	gp.stackguard0 = new.lo + _StackGuard // NOTE: might clobber a preempt request
   914  	gp.sched.sp = new.hi - used
   915  	gp.stktopsp += adjinfo.delta
   916  
   917  	// Adjust pointers in the new stack.
   918  	gentraceback(^uintptr(0), ^uintptr(0), 0, gp, 0, nil, 0x7fffffff, adjustframe, noescape(unsafe.Pointer(&adjinfo)), 0)
   919  
   920  	// free old stack
   921  	if stackPoisonCopy != 0 {
   922  		fillstack(old, 0xfc)
   923  	}
   924  	stackfree(old)
   925  }
   926  
   927  // round x up to a power of 2.
   928  func round2(x int32) int32 {
   929  	s := uint(0)
   930  	for 1<<s < x {
   931  		s++
   932  	}
   933  	return 1 << s
   934  }
   935  
   936  // Called from runtime·morestack when more stack is needed.
   937  // Allocate larger stack and relocate to new stack.
   938  // Stack growth is multiplicative, for constant amortized cost.
   939  //
   940  // g->atomicstatus will be Grunning or Gscanrunning upon entry.
   941  // If the scheduler is trying to stop this g, then it will set preemptStop.
   942  //
   943  // This must be nowritebarrierrec because it can be called as part of
   944  // stack growth from other nowritebarrierrec functions, but the
   945  // compiler doesn't check this.
   946  //
   947  //go:nowritebarrierrec
   948  func newstack() {
   949  	thisg := getg()
   950  	// TODO: double check all gp. shouldn't be getg().
   951  	if thisg.m.morebuf.g.ptr().stackguard0 == stackFork {
   952  		throw("stack growth after fork")
   953  	}
   954  	if thisg.m.morebuf.g.ptr() != thisg.m.curg {
   955  		print("runtime: newstack called from g=", hex(thisg.m.morebuf.g), "\n"+"\tm=", thisg.m, " m->curg=", thisg.m.curg, " m->g0=", thisg.m.g0, " m->gsignal=", thisg.m.gsignal, "\n")
   956  		morebuf := thisg.m.morebuf
   957  		traceback(morebuf.pc, morebuf.sp, morebuf.lr, morebuf.g.ptr())
   958  		throw("runtime: wrong goroutine in newstack")
   959  	}
   960  
   961  	gp := thisg.m.curg
   962  
   963  	if thisg.m.curg.throwsplit {
   964  		// Update syscallsp, syscallpc in case traceback uses them.
   965  		morebuf := thisg.m.morebuf
   966  		gp.syscallsp = morebuf.sp
   967  		gp.syscallpc = morebuf.pc
   968  		pcname, pcoff := "(unknown)", uintptr(0)
   969  		f := findfunc(gp.sched.pc)
   970  		if f.valid() {
   971  			pcname = funcname(f)
   972  			pcoff = gp.sched.pc - f.entry
   973  		}
   974  		print("runtime: newstack at ", pcname, "+", hex(pcoff),
   975  			" sp=", hex(gp.sched.sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n",
   976  			"\tmorebuf={pc:", hex(morebuf.pc), " sp:", hex(morebuf.sp), " lr:", hex(morebuf.lr), "}\n",
   977  			"\tsched={pc:", hex(gp.sched.pc), " sp:", hex(gp.sched.sp), " lr:", hex(gp.sched.lr), " ctxt:", gp.sched.ctxt, "}\n")
   978  
   979  		thisg.m.traceback = 2 // Include runtime frames
   980  		traceback(morebuf.pc, morebuf.sp, morebuf.lr, gp)
   981  		throw("runtime: stack split at bad time")
   982  	}
   983  
   984  	morebuf := thisg.m.morebuf
   985  	thisg.m.morebuf.pc = 0
   986  	thisg.m.morebuf.lr = 0
   987  	thisg.m.morebuf.sp = 0
   988  	thisg.m.morebuf.g = 0
   989  
   990  	// NOTE: stackguard0 may change underfoot, if another thread
   991  	// is about to try to preempt gp. Read it just once and use that same
   992  	// value now and below.
   993  	preempt := atomic.Loaduintptr(&gp.stackguard0) == stackPreempt
   994  
   995  	// Be conservative about where we preempt.
   996  	// We are interested in preempting user Go code, not runtime code.
   997  	// If we're holding locks, mallocing, or preemption is disabled, don't
   998  	// preempt.
   999  	// This check is very early in newstack so that even the status change
  1000  	// from Grunning to Gwaiting and back doesn't happen in this case.
  1001  	// That status change by itself can be viewed as a small preemption,
  1002  	// because the GC might change Gwaiting to Gscanwaiting, and then
  1003  	// this goroutine has to wait for the GC to finish before continuing.
  1004  	// If the GC is in some way dependent on this goroutine (for example,
  1005  	// it needs a lock held by the goroutine), that small preemption turns
  1006  	// into a real deadlock.
  1007  	if preempt {
  1008  		if !canPreemptM(thisg.m) {
  1009  			// Let the goroutine keep running for now.
  1010  			// gp->preempt is set, so it will be preempted next time.
  1011  			gp.stackguard0 = gp.stack.lo + _StackGuard
  1012  			gogo(&gp.sched) // never return
  1013  		}
  1014  	}
  1015  
  1016  	if gp.stack.lo == 0 {
  1017  		throw("missing stack in newstack")
  1018  	}
  1019  	sp := gp.sched.sp
  1020  	if sys.ArchFamily == sys.AMD64 || sys.ArchFamily == sys.I386 || sys.ArchFamily == sys.WASM {
  1021  		// The call to morestack cost a word.
  1022  		sp -= sys.PtrSize
  1023  	}
  1024  	if stackDebug >= 1 || sp < gp.stack.lo {
  1025  		print("runtime: newstack sp=", hex(sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n",
  1026  			"\tmorebuf={pc:", hex(morebuf.pc), " sp:", hex(morebuf.sp), " lr:", hex(morebuf.lr), "}\n",
  1027  			"\tsched={pc:", hex(gp.sched.pc), " sp:", hex(gp.sched.sp), " lr:", hex(gp.sched.lr), " ctxt:", gp.sched.ctxt, "}\n")
  1028  	}
  1029  	if sp < gp.stack.lo {
  1030  		print("runtime: gp=", gp, ", goid=", gp.goid, ", gp->status=", hex(readgstatus(gp)), "\n ")
  1031  		print("runtime: split stack overflow: ", hex(sp), " < ", hex(gp.stack.lo), "\n")
  1032  		throw("runtime: split stack overflow")
  1033  	}
  1034  
  1035  	if preempt {
  1036  		if gp == thisg.m.g0 {
  1037  			throw("runtime: preempt g0")
  1038  		}
  1039  		if thisg.m.p == 0 && thisg.m.locks == 0 {
  1040  			throw("runtime: g is running but p is not")
  1041  		}
  1042  
  1043  		if gp.preemptShrink {
  1044  			// We're at a synchronous safe point now, so
  1045  			// do the pending stack shrink.
  1046  			gp.preemptShrink = false
  1047  			shrinkstack(gp)
  1048  		}
  1049  
  1050  		if gp.preemptStop {
  1051  			preemptPark(gp) // never returns
  1052  		}
  1053  
  1054  		// Act like goroutine called runtime.Gosched.
  1055  		gopreempt_m(gp) // never return
  1056  	}
  1057  
  1058  	// Allocate a bigger segment and move the stack.
  1059  	oldsize := gp.stack.hi - gp.stack.lo
  1060  	newsize := oldsize * 2
  1061  
  1062  	// Make sure we grow at least as much as needed to fit the new frame.
  1063  	// (This is just an optimization - the caller of morestack will
  1064  	// recheck the bounds on return.)
  1065  	if f := findfunc(gp.sched.pc); f.valid() {
  1066  		max := uintptr(funcMaxSPDelta(f))
  1067  		needed := max + _StackGuard
  1068  		used := gp.stack.hi - gp.sched.sp
  1069  		for newsize-used < needed {
  1070  			newsize *= 2
  1071  		}
  1072  	}
  1073  
  1074  	if gp.stackguard0 == stackForceMove {
  1075  		// Forced stack movement used for debugging.
  1076  		// Don't double the stack (or we may quickly run out
  1077  		// if this is done repeatedly).
  1078  		newsize = oldsize
  1079  	}
  1080  
  1081  	if newsize > maxstacksize || newsize > maxstackceiling {
  1082  		if maxstacksize < maxstackceiling {
  1083  			print("runtime: goroutine stack exceeds ", maxstacksize, "-byte limit\n")
  1084  		} else {
  1085  			print("runtime: goroutine stack exceeds ", maxstackceiling, "-byte limit\n")
  1086  		}
  1087  		print("runtime: sp=", hex(sp), " stack=[", hex(gp.stack.lo), ", ", hex(gp.stack.hi), "]\n")
  1088  		throw("stack overflow")
  1089  	}
  1090  
  1091  	// The goroutine must be executing in order to call newstack,
  1092  	// so it must be Grunning (or Gscanrunning).
  1093  	casgstatus(gp, _Grunning, _Gcopystack)
  1094  
  1095  	// The concurrent GC will not scan the stack while we are doing the copy since
  1096  	// the gp is in a Gcopystack status.
  1097  	copystack(gp, newsize)
  1098  	if stackDebug >= 1 {
  1099  		print("stack grow done\n")
  1100  	}
  1101  	casgstatus(gp, _Gcopystack, _Grunning)
  1102  	gogo(&gp.sched)
  1103  }
  1104  
  1105  //go:nosplit
  1106  func nilfunc() {
  1107  	*(*uint8)(nil) = 0
  1108  }
  1109  
  1110  // adjust Gobuf as if it executed a call to fn
  1111  // and then stopped before the first instruction in fn.
  1112  func gostartcallfn(gobuf *gobuf, fv *funcval) {
  1113  	var fn unsafe.Pointer
  1114  	if fv != nil {
  1115  		fn = unsafe.Pointer(fv.fn)
  1116  	} else {
  1117  		fn = unsafe.Pointer(funcPC(nilfunc))
  1118  	}
  1119  	gostartcall(gobuf, fn, unsafe.Pointer(fv))
  1120  }
  1121  
  1122  // isShrinkStackSafe returns whether it's safe to attempt to shrink
  1123  // gp's stack. Shrinking the stack is only safe when we have precise
  1124  // pointer maps for all frames on the stack.
  1125  func isShrinkStackSafe(gp *g) bool {
  1126  	// We can't copy the stack if we're in a syscall.
  1127  	// The syscall might have pointers into the stack and
  1128  	// often we don't have precise pointer maps for the innermost
  1129  	// frames.
  1130  	//
  1131  	// We also can't copy the stack if we're at an asynchronous
  1132  	// safe-point because we don't have precise pointer maps for
  1133  	// all frames.
  1134  	//
  1135  	// We also can't *shrink* the stack in the window between the
  1136  	// goroutine calling gopark to park on a channel and
  1137  	// gp.activeStackChans being set.
  1138  	return gp.syscallsp == 0 && !gp.asyncSafePoint && atomic.Load8(&gp.parkingOnChan) == 0
  1139  }
  1140  
  1141  // Maybe shrink the stack being used by gp.
  1142  //
  1143  // gp must be stopped and we must own its stack. It may be in
  1144  // _Grunning, but only if this is our own user G.
  1145  func shrinkstack(gp *g) {
  1146  	if gp.stack.lo == 0 {
  1147  		throw("missing stack in shrinkstack")
  1148  	}
  1149  	if s := readgstatus(gp); s&_Gscan == 0 {
  1150  		// We don't own the stack via _Gscan. We could still
  1151  		// own it if this is our own user G and we're on the
  1152  		// system stack.
  1153  		if !(gp == getg().m.curg && getg() != getg().m.curg && s == _Grunning) {
  1154  			// We don't own the stack.
  1155  			throw("bad status in shrinkstack")
  1156  		}
  1157  	}
  1158  	if !isShrinkStackSafe(gp) {
  1159  		throw("shrinkstack at bad time")
  1160  	}
  1161  	// Check for self-shrinks while in a libcall. These may have
  1162  	// pointers into the stack disguised as uintptrs, but these
  1163  	// code paths should all be nosplit.
  1164  	if gp == getg().m.curg && gp.m.libcallsp != 0 {
  1165  		throw("shrinking stack in libcall")
  1166  	}
  1167  
  1168  	if debug.gcshrinkstackoff > 0 {
  1169  		return
  1170  	}
  1171  	f := findfunc(gp.startpc)
  1172  	if f.valid() && f.funcID == funcID_gcBgMarkWorker {
  1173  		// We're not allowed to shrink the gcBgMarkWorker
  1174  		// stack (see gcBgMarkWorker for explanation).
  1175  		return
  1176  	}
  1177  
  1178  	oldsize := gp.stack.hi - gp.stack.lo
  1179  	newsize := oldsize / 2
  1180  	// Don't shrink the allocation below the minimum-sized stack
  1181  	// allocation.
  1182  	if newsize < _FixedStack {
  1183  		return
  1184  	}
  1185  	// Compute how much of the stack is currently in use and only
  1186  	// shrink the stack if gp is using less than a quarter of its
  1187  	// current stack. The currently used stack includes everything
  1188  	// down to the SP plus the stack guard space that ensures
  1189  	// there's room for nosplit functions.
  1190  	avail := gp.stack.hi - gp.stack.lo
  1191  	if used := gp.stack.hi - gp.sched.sp + _StackLimit; used >= avail/4 {
  1192  		return
  1193  	}
  1194  
  1195  	if stackDebug > 0 {
  1196  		print("shrinking stack ", oldsize, "->", newsize, "\n")
  1197  	}
  1198  
  1199  	copystack(gp, newsize)
  1200  }
  1201  
  1202  // freeStackSpans frees unused stack spans at the end of GC.
  1203  func freeStackSpans() {
  1204  
  1205  	// Scan stack pools for empty stack spans.
  1206  	for order := range stackpool {
  1207  		lock(&stackpool[order].item.mu)
  1208  		list := &stackpool[order].item.span
  1209  		for s := list.first; s != nil; {
  1210  			next := s.next
  1211  			if s.allocCount == 0 {
  1212  				list.remove(s)
  1213  				s.manualFreeList = 0
  1214  				osStackFree(s)
  1215  				mheap_.freeManual(s, spanAllocStack)
  1216  			}
  1217  			s = next
  1218  		}
  1219  		unlock(&stackpool[order].item.mu)
  1220  	}
  1221  
  1222  	// Free large stack spans.
  1223  	lock(&stackLarge.lock)
  1224  	for i := range stackLarge.free {
  1225  		for s := stackLarge.free[i].first; s != nil; {
  1226  			next := s.next
  1227  			stackLarge.free[i].remove(s)
  1228  			osStackFree(s)
  1229  			mheap_.freeManual(s, spanAllocStack)
  1230  			s = next
  1231  		}
  1232  	}
  1233  	unlock(&stackLarge.lock)
  1234  }
  1235  
  1236  // getStackMap returns the locals and arguments live pointer maps, and
  1237  // stack object list for frame.
  1238  func getStackMap(frame *stkframe, cache *pcvalueCache, debug bool) (locals, args bitvector, objs []stackObjectRecord) {
  1239  	targetpc := frame.continpc
  1240  	if targetpc == 0 {
  1241  		// Frame is dead. Return empty bitvectors.
  1242  		return
  1243  	}
  1244  
  1245  	f := frame.fn
  1246  	pcdata := int32(-1)
  1247  	if targetpc != f.entry {
  1248  		// Back up to the CALL. If we're at the function entry
  1249  		// point, we want to use the entry map (-1), even if
  1250  		// the first instruction of the function changes the
  1251  		// stack map.
  1252  		targetpc--
  1253  		pcdata = pcdatavalue(f, _PCDATA_StackMapIndex, targetpc, cache)
  1254  	}
  1255  	if pcdata == -1 {
  1256  		// We do not have a valid pcdata value but there might be a
  1257  		// stackmap for this function. It is likely that we are looking
  1258  		// at the function prologue, assume so and hope for the best.
  1259  		pcdata = 0
  1260  	}
  1261  
  1262  	// Local variables.
  1263  	size := frame.varp - frame.sp
  1264  	var minsize uintptr
  1265  	switch sys.ArchFamily {
  1266  	case sys.ARM64:
  1267  		minsize = sys.StackAlign
  1268  	default:
  1269  		minsize = sys.MinFrameSize
  1270  	}
  1271  	if size > minsize {
  1272  		stackid := pcdata
  1273  		stkmap := (*stackmap)(funcdata(f, _FUNCDATA_LocalsPointerMaps))
  1274  		if stkmap == nil || stkmap.n <= 0 {
  1275  			print("runtime: frame ", funcname(f), " untyped locals ", hex(frame.varp-size), "+", hex(size), "\n")
  1276  			throw("missing stackmap")
  1277  		}
  1278  		// If nbit == 0, there's no work to do.
  1279  		if stkmap.nbit > 0 {
  1280  			if stackid < 0 || stackid >= stkmap.n {
  1281  				// don't know where we are
  1282  				print("runtime: pcdata is ", stackid, " and ", stkmap.n, " locals stack map entries for ", funcname(f), " (targetpc=", hex(targetpc), ")\n")
  1283  				throw("bad symbol table")
  1284  			}
  1285  			locals = stackmapdata(stkmap, stackid)
  1286  			if stackDebug >= 3 && debug {
  1287  				print("      locals ", stackid, "/", stkmap.n, " ", locals.n, " words ", locals.bytedata, "\n")
  1288  			}
  1289  		} else if stackDebug >= 3 && debug {
  1290  			print("      no locals to adjust\n")
  1291  		}
  1292  	}
  1293  
  1294  	// Arguments.
  1295  	if frame.arglen > 0 {
  1296  		if frame.argmap != nil {
  1297  			// argmap is set when the function is reflect.makeFuncStub or reflect.methodValueCall.
  1298  			// In this case, arglen specifies how much of the args section is actually live.
  1299  			// (It could be either all the args + results, or just the args.)
  1300  			args = *frame.argmap
  1301  			n := int32(frame.arglen / sys.PtrSize)
  1302  			if n < args.n {
  1303  				args.n = n // Don't use more of the arguments than arglen.
  1304  			}
  1305  		} else {
  1306  			stackmap := (*stackmap)(funcdata(f, _FUNCDATA_ArgsPointerMaps))
  1307  			if stackmap == nil || stackmap.n <= 0 {
  1308  				print("runtime: frame ", funcname(f), " untyped args ", hex(frame.argp), "+", hex(frame.arglen), "\n")
  1309  				throw("missing stackmap")
  1310  			}
  1311  			if pcdata < 0 || pcdata >= stackmap.n {
  1312  				// don't know where we are
  1313  				print("runtime: pcdata is ", pcdata, " and ", stackmap.n, " args stack map entries for ", funcname(f), " (targetpc=", hex(targetpc), ")\n")
  1314  				throw("bad symbol table")
  1315  			}
  1316  			if stackmap.nbit > 0 {
  1317  				args = stackmapdata(stackmap, pcdata)
  1318  			}
  1319  		}
  1320  	}
  1321  
  1322  	// stack objects.
  1323  	if GOARCH == "amd64" && unsafe.Sizeof(abi.RegArgs{}) > 0 && frame.argmap != nil {
  1324  		// argmap is set when the function is reflect.makeFuncStub or reflect.methodValueCall.
  1325  		// We don't actually use argmap in this case, but we need to fake the stack object
  1326  		// record for these frames which contain an internal/abi.RegArgs at a hard-coded offset
  1327  		// on amd64.
  1328  		objs = methodValueCallFrameObjs
  1329  	} else {
  1330  		p := funcdata(f, _FUNCDATA_StackObjects)
  1331  		if p != nil {
  1332  			n := *(*uintptr)(p)
  1333  			p = add(p, sys.PtrSize)
  1334  			*(*slice)(unsafe.Pointer(&objs)) = slice{array: noescape(p), len: int(n), cap: int(n)}
  1335  			// Note: the noescape above is needed to keep
  1336  			// getStackMap from "leaking param content:
  1337  			// frame".  That leak propagates up to getgcmask, then
  1338  			// GCMask, then verifyGCInfo, which converts the stack
  1339  			// gcinfo tests into heap gcinfo tests :(
  1340  		}
  1341  	}
  1342  
  1343  	return
  1344  }
  1345  
  1346  var (
  1347  	abiRegArgsEface          interface{} = abi.RegArgs{}
  1348  	abiRegArgsType           *_type      = efaceOf(&abiRegArgsEface)._type
  1349  	methodValueCallFrameObjs             = []stackObjectRecord{
  1350  		{
  1351  			off:      -int32(alignUp(abiRegArgsType.size, 8)), // It's always the highest address local.
  1352  			size:     int32(abiRegArgsType.size),
  1353  			_ptrdata: int32(abiRegArgsType.ptrdata),
  1354  			gcdata:   abiRegArgsType.gcdata,
  1355  		},
  1356  	}
  1357  )
  1358  
  1359  func init() {
  1360  	if abiRegArgsType.kind&kindGCProg != 0 {
  1361  		throw("abiRegArgsType needs GC Prog, update methodValueCallFrameObjs")
  1362  	}
  1363  }
  1364  
  1365  // A stackObjectRecord is generated by the compiler for each stack object in a stack frame.
  1366  // This record must match the generator code in cmd/compile/internal/liveness/plive.go:emitStackObjects.
  1367  type stackObjectRecord struct {
  1368  	// offset in frame
  1369  	// if negative, offset from varp
  1370  	// if non-negative, offset from argp
  1371  	off      int32
  1372  	size     int32
  1373  	_ptrdata int32 // ptrdata, or -ptrdata is GC prog is used
  1374  	gcdata   *byte // pointer map or GC prog of the type
  1375  }
  1376  
  1377  func (r *stackObjectRecord) useGCProg() bool {
  1378  	return r._ptrdata < 0
  1379  }
  1380  
  1381  func (r *stackObjectRecord) ptrdata() uintptr {
  1382  	x := r._ptrdata
  1383  	if x < 0 {
  1384  		return uintptr(-x)
  1385  	}
  1386  	return uintptr(x)
  1387  }
  1388  
  1389  // This is exported as ABI0 via linkname so obj can call it.
  1390  //
  1391  //go:nosplit
  1392  //go:linkname morestackc
  1393  func morestackc() {
  1394  	throw("attempt to execute system stack code on user stack")
  1395  }
  1396  

View as plain text